A method and device for spatiotemporal prediction of sea ice density

Abstract

The present invention provides a method and device for spatiotemporal prediction of sea ice density. The method comprises obtaining a seasonal characteristic map and a trend characteristic map of the target area based on the acquired measured sea ice density data of the target area in a preset historical period and a spatiotemporal prediction model of sea ice density, and obtaining sea ice density prediction data of the target area in a future period based on the seasonal characteristic map and the trend characteristic map. The present invention predicts sea ice density using a constructed spatiotemporal prediction model of sea ice density, and does not use a commonly used cyclic time network, thereby overcoming the problem of time dependence, improving the accuracy of sea ice prediction, and reducing model complexity and computational overhead.

Description

Technical Field

[0001] One or more embodiments of the present specification relate to the technical field of ice condition prediction applications, and in particular, to a method and device for spatiotemporal prediction of sea ice density. Background Art

[0002] Sea ice concentration (SIC) refers to the proportion of sea ice coverage in a certain area to the total area of ​​the area, reflecting the density of sea ice distribution in a specific area. At present, sea ice concentration can be predicted based on physical laws such as sea ice dynamics, thermodynamics and thickness distribution, but this method requires complex physical parameters and has high computational costs, and the rapid changes in sea ice will lead to inaccurate prediction results. Therefore, establishing a deep learning model to predict sea ice prediction is another effective method. Because it is necessary to process information in the time and space dimensions at the same time, traditional time recurrent networks or convolutional long short-term memory networks and other time-related recurrent neural networks are usually used. When the time series is too long, the gradient disappears or explodes, resulting in the inability to obtain prediction results. In order to improve the accuracy of sea ice density prediction, the time series required in the prediction process will be very long. Summary of the invention

[0003] The present invention describes a method and device for spatiotemporal prediction of sea ice density, which can solve the above technical problems.

[0004] According to a first aspect, a method for spatiotemporal prediction of sea ice density is provided. The method comprises:

[0005] Obtaining measured data of sea ice density in a target area during a preset historical period, wherein the spatiotemporal prediction model of sea ice density is constructed based on a global sparse attention mechanism and a convolutional neural network and is trained using the historical measured data of sea ice density in the target area;

[0006] According to the measured data of sea ice density of the target area in a preset historical period and the spatiotemporal prediction model of sea ice density, a seasonal characteristic map of the target area and a trend characteristic map of the target area are obtained, wherein the seasonal characteristic map is used to represent the seasonal characteristics of the sea ice in the target area changing with the seasonal law in the preset historical period, and the trend characteristic map is used to represent the spatial variation trend of the sea ice in the target area in the preset historical period;

[0007] Based on the seasonal characteristic map and the trend characteristic map, the sea ice density prediction data of the target area in the future period is obtained.

[0008] In some embodiments, the spatiotemporal prediction model of sea ice density includes a seasonal network, and the seasonal network is used to extract a trend characteristic graph of sea ice changing with seasonal patterns;

[0009] According to the measured sea ice density data of the target area in a preset historical period and the spatiotemporal prediction model of sea ice density, a seasonal characteristic map of the target area is obtained, which specifically includes:

[0010] Extracting a seasonal component from the measured sea ice density data, wherein the seasonal component is used to represent a seasonal regular part of the measured sea ice density data;

[0011] The seasonal component is input into the seasonal network to obtain the seasonal characteristic map.

[0012] In some embodiments, the sea ice density spatiotemporal prediction model includes a trend network, and the trend network is used to obtain a characteristic map of the spatial variation trend of sea ice;

[0013] The obtaining, according to the at least one sea ice density time data series and the sea ice density spatiotemporal prediction model, a trend characteristic map of the target area specifically comprises:

[0014] Extracting a trend component from the first sea ice concentration time data series, wherein the trend component is used to represent a trend change portion in the sea ice concentration time data series;

[0015] The seasonal component is input into the seasonal network to obtain the seasonal characteristic map.

[0016] In some more specific embodiments, the season network includes an encoder, a global sparse attention mechanism GSAM, and a decoder;

[0017] The encoder comprises at least one convolutional layer for encoding high-dimensional frames in the seasonal component into a low-dimensional latent space;

[0018] The GSAM comprises at least one spatiotemporal attention mechanism for learning latent spatiotemporal information in the seasonal component from a low-dimensional latent space;

[0019] The decoder comprises at least one transposed convolutional layer for decoding the potential spatiotemporal information in the seasonal component to obtain the seasonal feature map;

[0020] The step of inputting the seasonal component into the seasonal network to obtain the seasonal characteristic graph specifically includes:

[0021] Inputting the seasonal component into the encoder for encoding to obtain a first feature map of a low-dimensional latent space;

[0022] Inputting the first feature map into the GSAM to obtain a second feature map containing potential spatiotemporal information;

[0023] The second characteristic map is input into the decoder for decoding to obtain the seasonal characteristic map.

[0024] In some more specific embodiments, the GSAM includes a channel attention mechanism and a spatial attention mechanism, the channel attention mechanism includes a sparse multilayer perceptron sMLP for compressing the spatial dimension of the first feature map, and the spatial attention mechanism includes at least one convolutional layer for fusing spatial information of the feature map output by the channel attention mechanism;

[0025] The step of inputting the first characteristic graph into the GSAM to obtain a second characteristic graph output by the GSAM specifically includes:

[0026] Inputting the first feature map into the channel attention mechanism for spatial dimension compression to obtain a first convolution feature map output by the channel attention mechanism;

[0027] Multiplying the first feature map and the first convolutional feature map to obtain a first input feature map;

[0028] Inputting the first input feature map into the spatial attention mechanism to perform spatial information fusion to obtain a second convolution feature map;

[0029] The second convolutional feature map is multiplied by the first input feature map to obtain the second feature map.

[0030] In some embodiments, the trend network includes at least one convolutional layer for extracting trend features in the trend component;

[0031] The step of inputting the trend component into the trend network to obtain the trend characteristic graph specifically includes:

[0032] Using the trend component Get the trend characteristics Figure X TN ;

[0033] in, is the trend component X t , is the trend component obtained by the i-th round of convolution, σ is the nonlinear activation function, GroupNorm is the normalization layer, and N T is the number of convolutional layers in the trend network, and Conv2d is the convolutional layer.

[0034] In some embodiments, predicting the sea ice density prediction data of the target area in the second future period based on the seasonal characteristic map and the trend characteristic map specifically includes:

[0035] The seasonal characteristic graph and the trend characteristic graph are added together to obtain the sea ice density prediction data of the target area in the second future period.

[0036] According to a second aspect, a device for spatiotemporal prediction of sea ice density is provided. The device comprises:

[0037] A construction module is used to obtain measured data of sea ice density in a target area in a preset historical period, wherein the spatiotemporal prediction model of sea ice density is constructed based on a global sparse attention mechanism and a convolutional neural network, and is trained using the historical measured data of sea ice density in the target area;

[0038] A processing module, for obtaining a seasonal characteristic map and a trend characteristic map of the target area according to the measured sea ice density data of the target area in a preset historical period and the spatiotemporal prediction model of the sea ice density, wherein the seasonal characteristic map is used to represent the seasonal characteristics of the sea ice in the target area changing with the seasonal law in the preset historical period, and the trend characteristic map is used to represent the spatial variation trend of the sea ice in the target area in the preset historical period;

[0039] A fusion module is used to obtain the sea ice density prediction data of the target area in the future period based on the seasonal characteristic map and the trend characteristic map.

[0040] According to a third aspect, a computer storage medium is provided, wherein a computer program is stored on the computer-readable storage medium, and when the computer program is executed by one or more processors, the method for spatiotemporal prediction of sea ice density as described in any one of the above embodiments is implemented.

[0041] According to a fourth aspect, an electronic device is provided, comprising a memory and one or more processors, wherein the memory stores a computer program, and when the computer program is executed by the one or more processors, the method for spatiotemporal prediction of sea ice density as described in any one of the above embodiments is implemented.

[0042] Compared with the prior art, the method and device for spatiotemporal prediction of sea ice density provided in the embodiments of the present invention use the seasonal characteristic graph and trend characteristic graph obtained by the constructed spatiotemporal prediction model of sea ice density to predict the sea ice density, which is in line with the obvious seasonal fluctuation and time trend change properties of sea ice density. It does not adopt traditional time-related recurrent neural networks such as time recurrent networks or convolutional long short-term memory networks, thus overcoming the problem of model dependence on time, improving the accuracy of sea ice prediction, and reducing model complexity and computational overhead. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0044] Figure 1 A schematic diagram showing a framework of a spatiotemporal prediction model for sea ice density provided in an embodiment of this specification;

[0045] Figure 2 A schematic flow chart showing a method for spatiotemporal prediction of sea ice density provided in an embodiment of this specification;

[0046] Figure 3 A schematic diagram of SSTNet input data provided in an embodiment of this specification is shown;

[0047] Figure 4 A schematic diagram showing the overall process of SSTNet provided in the embodiments of this specification;

[0048] Figure 5 A schematic diagram of the framework of the GSAM provided in the embodiments of this specification is shown;

[0049] Figure 6 A schematic diagram showing the MLP and Sprase MLP provided in the embodiments of this specification;

[0050] Figure 7 A schematic diagram of Arctic sea ice density in March and September provided in an embodiment of this specification is shown;

[0051] Figure 8 A performance diagram showing a model test provided in an embodiment of this specification;

[0052] Fig. 9 A schematic diagram showing observation results and prediction results of the sea ice edge during the melting season provided by an embodiment of this specification;

[0053] Fig.10 The SSTNet provided in the embodiment of this specification is compared with other methods in terms of three indicators (a) MAE, (b) MAPE and (c) R 2 Schematic diagram of quantitative time step comparison on ;

[0054] Fig.11 A schematic diagram showing the comparison of prediction performance of SSTNet from March 12 to March 21 provided in the embodiments of this specification with other 9 methods;

[0055] Fig.12A schematic diagram showing the residual distribution of the sea ice density predicted by each model and the observed sea ice density from March 12 to March 21 provided in the embodiments of this specification is shown;

[0056] Fig.13 The results show that the SSTNet provided by the embodiments of this specification from September 8 to September 17 is comparable to other methods in terms of three indicators (a) MAE, (b) MAPE and (c) R 2 Schematic diagram of quantitative time step comparison on ;

[0057] Fig.14 A schematic diagram showing the prediction performance comparison between SSTNet and other 9 methods from September 8 to September 17 provided in the embodiments of this specification is shown;

[0058] Fig.15 A schematic diagram showing the residual distribution of the sea ice density predicted by each model and the observed sea ice density from September 8 to September 17 provided in the embodiments of this specification is shown;

[0059] Fig.16 A schematic diagram of the structure of a device for spatiotemporal prediction of sea ice density provided in an embodiment of this specification is shown. DETAILED DESCRIPTION

[0060] The solution provided in this specification is described below in conjunction with the accompanying drawings.

[0061] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0062] In the description of the embodiments of the present application, words such as "exemplary", "for example" or "for example" are used to indicate examples, illustrations or descriptions. Any embodiment or design described as "exemplary", "for example" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary", "for example" or "for example" is intended to present related concepts in a concrete way.

[0063] In the description of the embodiments of the present application, the term "and / or" is merely a description of the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B may represent: A exists alone, B exists alone, and A and B exist at the same time. In addition, unless otherwise specified, the term "plurality" means two or more.

[0064] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. The terms "include", "comprises", "has" and their variations all mean "including but not limited to", unless otherwise specifically emphasized.

[0065] It should be noted that in the time recurrent network RNNs or the Convolutional LSTM Network (ConvLSTM) that combines convolutional neural networks and long short-term memory recurrent neural networks, in order to process data with sequential properties, each element in these data not only contains its own feature information, but may also depend on other elements in the sequence, especially previous elements. Time dependency means that the output of the current time step depends not only on the input of the current time step, but also on the information of the previous time step. When the network tries to capture the dependencies between elements that are far apart in the sequence, gradients are prone to vanishing or exploding, resulting in the failure of the model to converge. For short-term predictions of sea ice density, it is necessary to input time series of 5 days, 10 days or even longer, and also consider information from multiple historical moments, which will undoubtedly aggravate the defects of the model.

[0066] The average value of sea ice concentration (SIC) has obvious seasonal fluctuations, indicating that SIC has obvious trend and seasonality. In recent years, many works have studied how to use convolutional neural networks and recurrent neural networks to process SIC data, and some works have discussed how to effectively use periodic information to improve the accuracy of predictions. However, deep neural networks are plagued by gradient vanishing and error accumulation, and it is difficult to capture long-term dependencies in time series data. Therefore, how to overcome the difficulty of neural networks in capturing long-term dependencies in time series data, and how to use the trend and seasonality of SIC for modeling, and extracting spatiotemporal signals with trend and seasonality for sea ice concentration prediction is a technical problem that needs to be solved urgently.

[0067] In view of this, the present application proposes a method and device for spatiotemporal prediction of sea ice density, and establishes a spatiotemporal prediction model of sea ice density based on SSTNet (Sparse Spatial-Temporal Network).

[0068] like Figure 1 As shown in the SSTNet framework diagram, the spatiotemporal prediction model of sea ice density is divided into two parts: seasonal network and trend network. The trend network includes multiple 1×1 convolutional layers. The sea ice evolution trend is inferred through the trend component to obtain a trend feature map.

[0069] The seasonal network includes an encoder, a GSAM (Global Sparse Attention Module) module, and a decoder, which infers the seasonal characteristics of sea ice evolution through seasonal components. The seasonal network uses an encoder to encode high-dimensional past frames into a low-dimensional latent space. The GSAM low-dimensional latent space learns the potential spatiotemporal information including spatial dependencies and temporal changes. The decoder decodes the potential spatiotemporal information to obtain a seasonal feature map, and fuses the seasonal feature map with the trend feature map to obtain sea ice density prediction data including spatiotemporal predictions.

[0070] like Figure 4 (a) The overall flow chart of SSTNet shows that the sea ice density dataset is decomposed into trend components X as input data. t and seasonal component X s , using two different networks to analyze the seasonal component X s and trend component X t To make predictions, we need to consider the seasonal component X s The input seasonal network predicts the seasonal feature map. The seasonal network includes an encoder, GSAM and a decoder.

[0071] The trend component X t Input the trend network to predict the trend feature map. The trend network is a convolutional network. The seasonal feature map and the trend feature map are fused to obtain the prediction result.

[0072] Figure 4 (b) The input data of the encoder in the seasonal network is the seasonal component, which includes N sub-modules, each of which includes a convolutional layer, a normalization layer and an activation function.

[0073] The seasonal component enters the submodule for iterative calculation, and the output feature map is used as the input data of GSAM. Figure 4 (c) GSAM module. In terms of processing order, the GSAM module includes convolutional layer, channel attention mechanism, spatial attention mechanism and convolutional layer in sequence. The feature map output by the encoder is processed by these modules in sequence, and the obtained feature map is used as the input data of the decoder.

[0074] Figure 4 (d) The decoder consists of N submodules, each of which includes a transposed convolution layer, a normalization layer, and an activation function. The feature map enters each submodule for iterative calculation and outputs a seasonal feature map.

[0075] Figure 2 A flow chart of a method for spatiotemporal prediction of sea ice density provided in an embodiment of this specification is shown, and the method comprises the following steps:

[0076] Step S110, obtaining measured data on sea ice density in a target area in a preset historical period, wherein a spatiotemporal prediction model for sea ice density is constructed based on a global sparse attention mechanism and a convolutional neural network, and is trained using the historical measured data on sea ice density in the target area.

[0077] In some embodiments, the sea ice concentration data of a preset historical period and the sea ice concentration prediction data of a first future period are combined according to time to form sea ice concentration time data;

[0078] Among them, the sea ice density prediction data for the first future period is the output result obtained by inputting the measured sea ice density data for a preset historical period into the trained sea ice density spatiotemporal prediction model.

[0079] The spatiotemporal prediction model of sea ice density is built based on the global sparse attention mechanism and convolutional neural network.

[0080] A sliding time window is used to slide on the sea ice concentration time data to obtain at least one sea ice concentration time data series.

[0081] The sea ice concentration time data can be slid based on the set sliding window length, and the step length of each sliding can be a time period.

[0082] For example, if the sea ice density for the next 10 days is to be predicted, the sea ice density data for the previous 8 days' historical period is needed. The next 10 days and the previous 8 days' historical period are taken as a time series, the length of the sliding window is set to 8, and the sliding step is 1. Then, when predicting the sea ice density for the first day, the sliding window range is the previous 8-day historical period, and the data set corresponding to the sliding window range is the sea ice density for the previous 8-day historical period. When predicting the sea ice density for the second day, the sliding window range is the previous 8-day historical period, which is the historical period from the 2nd day to the 8th day and the predicted first day. The data set corresponding to the sliding window range is the sea ice density for the previous 2nd day to the 8th day and the predicted sea ice density for the predicted first day.

[0083] The above approach meets the needs of sea ice density prediction for information at multiple historical moments and information on the current prediction time series, while limiting the length of the input time series, overcoming the time dependency problem, and avoiding the gradient vanishing or exploding phenomenon in deep learning models.

[0084] In addition, according to a specific implementation, the data set corresponding to the sliding window range can be divided into input data for prediction and data for supervised training.

[0085] The sea ice density data of a preset historical period is processed to obtain a historical sea ice density data sequence, and the historical sea ice density data sequence is input into a trained sea ice density spatiotemporal prediction model to obtain sea ice density prediction data for the first future period.

[0086] Specifically, the earliest time period of the first future time period and the latest time period of the preset historical time period are separated by one cycle.

[0087] At the same time, the historical measured data of sea ice density in the historical period and the predicted data of sea ice density in the future period are taken as a time data series, wherein the historical period can be the historical period before the earliest period in the future period.

[0088] The sliding window slides in the time data series based on the set sliding window length. The step size of each sliding can be a prediction cycle. The sliding window can include historical measured data for all or part of the historical period and all or part of the predicted data for the future period.

[0089] Each time a cycle is predicted, all data in the sea ice density time data series corresponding to each sliding window range are obtained. When predicting the next cycle, the sliding window is moved according to the step size to keep the length of the time data series input into the model each time controllable.

[0090] Step S120, based on the measured data of sea ice density in the target area in a preset historical period and the spatiotemporal prediction model of sea ice density, a seasonal characteristic map and a trend characteristic map of the target area are obtained. The seasonal characteristic map is used to represent the seasonal characteristics of the sea ice in the target area that changes with the seasonal laws in the preset historical period, and the trend characteristic map is used to represent the spatial change trend of the sea ice in the target area in the preset historical period.

[0091] Specifically, seasonal characteristics and trend characteristics respectively describe the change patterns of data at different time scales. The model in this application is non-cyclic, that is, within a forecast period, no information from the previous or previous forecast periods is introduced.

[0092] Trend characteristics refer to short-term trends, and seasonal characteristics are also short-term seasonal characteristics. The relationship between the two is a mutually influential relationship. Seasonal characteristics may mask trend characteristics, so the two are decomposed and predicted separately.

[0093] In some embodiments, the sea ice density data set is average pooled to obtain a trend component, and the trend component is removed from the sea ice density data set to obtain a seasonal component.

[0094] Specifically, for a sea ice density dataset of length L AvgPool(·) is used for moving average, that is, the corresponding elements of the two pictures are averaged to obtain X t ,

[0095] X t =AvgPool(X)

[0096] X s =XX t

[0097] X s represents the seasonal component and X t Represents the trend component.

[0098] In some embodiments, the spatiotemporal prediction model of sea ice density includes a seasonal network, and the seasonal network is used to extract a trend characteristic map of sea ice changing with seasonal laws;

[0099] Based on the measured data of sea ice density in the target area during the preset historical period and the spatiotemporal prediction model of sea ice density, the seasonal characteristic map of the target area is obtained, including:

[0100] Extracting seasonal components from measured sea ice density data, where the seasonal components are used to represent the seasonal regularity of the measured sea ice density data;

[0101] The seasonal component is input into the seasonal network to obtain the seasonal feature map.

[0102] In some other embodiments, obtaining a seasonal characteristic map of a target area based on at least one sea ice density time data series and a sea ice density spatiotemporal prediction model specifically includes:

[0103] Extracting a seasonal component from the first sea ice concentration time data series, the seasonal component is used to represent a seasonal regular part in the sea ice concentration time data series;

[0104] The seasonal component is input into the seasonal network to obtain the seasonal feature map.

[0105] In some embodiments, the sea ice density spatiotemporal prediction model includes a trend network, and the trend network is used to obtain a characteristic map of the spatial variation trend of sea ice;

[0106] Based on the measured data of sea ice density in the target area during the preset historical period and the spatiotemporal prediction model of sea ice density, the trend characteristic map of the target area is obtained, including:

[0107] Extracting a trend component from the measured sea ice density data, the trend component is used to represent the trend change part in the measured sea ice density data;

[0108] The seasonal component is input into the seasonal network to obtain the seasonal feature map.

[0109] In some more specific embodiments, the season network includes an encoder, a global sparse attention mechanism GSAM, and a decoder.

[0110] The encoder includes at least one convolutional layer for encoding a high-dimensional frame of the sea ice density dataset into a low-dimensional latent space, and the GSAM includes at least one spatiotemporal attention mechanism for learning latent spatiotemporal information in the sea ice density dataset from the low-dimensional latent space;

[0111] The decoder comprises at least one transposed convolutional layer for decoding the potential spatiotemporal information in the seasonal component to obtain a seasonal feature map;

[0112] Accordingly, the implementation of this step may also include the following steps:

[0113] 121. Input the seasonal component into the encoder for encoding to obtain a first feature map of the low-dimensional latent space.

[0114] 122. Input the first feature map into GSAM to obtain a second feature map containing potential spatiotemporal information.

[0115] Further, in some more specific embodiments, GSAM includes a channel attention mechanism and a spatial attention mechanism. The channel attention mechanism includes a sparse multilayer perceptron sMLP for compressing the spatial dimension of the first feature map. The spatial attention mechanism includes at least one convolutional layer for performing spatial information fusion on the feature map output by the channel attention mechanism.

[0116] Accordingly, the implementation of step 122 may further include:

[0117] 1221. Input the first feature map into the channel attention mechanism for spatial dimension compression to obtain the first convolution feature map output by the channel attention mechanism.

[0118] 1222. Multiply the first feature map and the first convolution feature map to obtain a first input feature map.

[0119] 1223. Input the first input feature map into the spatial attention mechanism for spatial information fusion to obtain a second convolution feature map.

[0120] 1224. Multiply the second convolution feature map and the first input feature map to obtain a second feature map.

[0121] 123. Input the second feature map into a decoder for decoding to obtain a seasonal feature map.

[0122] In some other embodiments, the trend network includes at least one convolution layer for extracting trend features from trend components. Accordingly, the trend components are input into the trend network to obtain a trend feature graph, which specifically includes:

[0123] use Get trend characteristics Figure X TN ;

[0124] in, is the trend component X t , is the trend component obtained by the i-th round of convolution, σ is the nonlinear activation function, GroupNorm is the normalization layer, and N T is the number of convolutional layers in the trend network, and Conv2d is the convolutional layer.

[0125] Step S130, based on the seasonal characteristic graph and the trend characteristic graph, obtain the sea ice density prediction data of the target area in the future period.

[0126] In some embodiments, the implementation of this step may include:

[0127] The seasonal characteristic map and the trend characteristic map are added together to obtain the predicted data of sea ice density in the target area in the future period.

[0128] In some embodiments, based on the seasonal characteristic map and the trend characteristic map, sea ice density prediction data for the target area in a second future period is predicted.

[0129] The end time of the first future time period is earlier than the start time of the second future time period.

[0130] Based on the sea ice density prediction method in the above embodiment, a lightweight non-cyclic spatiotemporal prediction model SSTNet is constructed to predict the temporal and spatial distribution of sea ice density. The model is divided into two parts, season and trend, based on the STL decomposition idea, and the GSAM module is introduced, so that SSTNet can not only extract seasonal signals and trend information in the data, but also capture the spatiotemporal dependencies in the data, and end-to-end predict the daily spatiotemporal distribution of sea ice density in the next few days.

[0131] This embodiment uses a seasonal network and a trend network to predict sea ice density, which is consistent with the obvious seasonal fluctuation and time trend change properties of sea ice density. The commonly used cyclic time network is not used, which overcomes the problem of model dependence on time. At the same time, the historical measured data of sea ice density and the predicted data of sea ice density are used for prediction, which improves the accuracy of sea ice prediction and reduces the model complexity and computational overhead. In the context of the gradual warming of the global climate, the prediction of sea ice density has important practical significance, which not only provides a basis for subsequent research, but also provides safety for ship navigation.

[0132] Next, further combining the attached Figure 3 , Attachment Figures 5 and 6, describing a sea ice density prediction method according to an embodiment of the present invention, including:

[0133] Step 1: Obtain the daily spatial distribution data of sea ice density. The daily data records the value of the sea ice density element of each coordinate grid within the target area.

[0134] Specifically, the time resolution of the daily spatial distribution data of sea ice density is 1 day, the number of time frames is T days, and the sea ice density data with a grid data size of H×W within the target area are combined into a sea ice density dataset in the form of T×H×W.

[0135] Step 2: Construct a spatiotemporal prediction model for sea ice density based on the SSTNet spatiotemporal prediction network model.

[0136] Step 2.1: Decompose the sea ice concentration dataset into trend component and seasonal component according to the time series variation law.

[0137] Step 2.2: Use two different networks to analyze the seasonal component X s and trend component X t Make predictions.

[0138] like Figure 3 (a) shows the sea ice density dataset B is the batch number, T is the number of frames, C is the number of channels, and H×W is the frame image size. The sea ice density dataset G is decomposed into trend component and seasonal component. Both trend component and seasonal component are reshaped into tensors of shape (B×T)×C×H×W and then input into trend network and seasonal network respectively.

[0139] like Figure 3 As shown in (b), in GSAM, the hidden representation of the encoder output is reshaped into a tensor of shape B×(T×C)×H×W, and multi-frame level features are stacked along the time axis. The channel attention module and spatial attention module designed based on the convolutional network are learned from the multi-frame level features to capture the inherent spatiotemporal evolution characteristics of the sea ice density dataset.

[0140] In order to predict the trend component, construct T A trend network with a 1×1 convolutional layer and a fixed hidden size of n T , the evolution trend can be directly inferred from the trend component, as follows:

[0141]

[0142] Where σ is a nonlinear activation function. This embodiment uses the LeakyReLU activation function. GroupNorm is a normalization layer. T is the trend component of the input.

[0143] In order to predict the seasonal component, a seasonal network with an encoder-decoder structure is constructed, which includes an encoder, a GSAM module, and a decoder. S The encoder has N convolutional layers and S The decoder has a transposed convolutional layer. The hidden size of both the encoder and decoder is n. S The hidden representation f in the encoder can be expressed as follows:

[0144]

[0145] Where X s0 is the seasonal component of the input. For every two convolutional layers, downsampling is performed once, that is, if downsampling is required, a convolutional layer with a stride of 2 is used. If downsampling is not required, a convolutional layer with a stride of 1 is used to ensure that the image size remains unchanged.

[0146] The hidden representation f in the decoder -1 It can be expressed as follows:

[0147]

[0148] ConvTranspose2d is a transposed convolutional layer. Consistent with the encoder, upsampling is performed once for every two convolutional layers, that is, if upsampling is required, a transposed convolutional layer is used. Otherwise, a convolutional layer with a stride of 1 is used.

[0149] GSAM module is composed of N C The spatiotemporal attention modules are composed of a fixed hidden size of n. C The hidden representation of the GSAM module can be expressed as follows:

[0150]

[0151] in The spatial features are maintained by introducing a residual connection from the first layer of the encoder to the last layer of the decoder as the output features of the encoder.

[0152] In order to amplify cross-dimensional interactions and reduce computational costs, this embodiment proposes a GSAM attention mechanism, such as Figure 5 As shown, the GSAM attention mechanism is divided into Figure 5 (a) Channel attention module and Figure 5 (b) The spatial attention module consists of two parts. Figure 5(a) The channel attention module compresses the feature map in the spatial dimension to generate a channel attention map. For a picture, the channel attention module focuses on which content in the picture is important. Use 3D permutation to retain three-dimensional information In order to effectively calculate the channel attention, the Sparse Multilayer Perceptron (sMLP) is used to compress the spatial dimension of the input feature map to amplify the cross-dimensional channel-space dependency. For example, the C×W×H frame image is converted to a (H×C)×W frame image. After using MLP to process the W dimension, the output feature map of the module is multiplied by the input feature map and permuted into the features of the input dimension. The C×W×H frame image is converted to a (W×C)×H frame image. After using MLP to process the H dimension, the output feature map of the module is multiplied by the input feature map and permuted into the features of the input dimension. After all the features are fused, they are input into the spatial attention module.

[0153] Figure 5 (b) The spatial attention mechanism compresses the channel. The spatial attention module focuses on the most informative position on a picture, which is a supplement to the channel attention. The feature map output by the channel attention module is used as the input feature map of this module. Two convolutional layers are used to fuse spatial information. The reduction ratio is set to r to compress the channel and reduce the dimension to C / r channels. The feature map is multiplied with the input feature map of the module to obtain a feature map. For example, after the C×W×H frame image is processed by a 7×7 convolutional layer, the dimension is reduced through the compression channel to obtain a feature of dimension C / r×W×H. After the feature of dimension C / r×W×H is processed by a 7×7 convolutional layer, a feature map of dimension C×W×H is obtained.

[0154] In the channel attention module, this embodiment replaces the traditional MLP with sMLP, which reduces the number of parameters to avoid overfitting. At the same time, it reduces the computational complexity, especially when the number of input units is large, to achieve multi-stage processing in the pyramid structure.

[0155] In the Sparse Multilayer Perceptron (sMLP), sparse connections and weight sharing are used to achieve the design goals. Figure 6 Units in the MLP shown in (a) can interact with all other units. Figure 6 As shown in (b), the units in the sMLP only interact directly with the units on the same row or column, not with all other units. In addition, all rows and all columns can share the same projection weights, respectively. The units in the sMLP consist of three paths. In addition to the identity mapping shown in the middle, the other two paths are along the horizontal and vertical directions, respectively. Therefore, using sMLP can not only efficiently aggregate spatial dimensions, but also reduce the amount of computation.

[0156] Specifically, the training method of the sea ice density spatiotemporal prediction model is as follows:

[0157] The sea ice density dataset is split into a training set, a validation set, and a test set.

[0158] End-to-end training is adopted, and the number of layers of the trend network is set to L1, the number of hidden sizes is set to H1, the number of layers of the encoder encoder and decoder decoder in the season network is set to L2, and the number of hidden sizes is set to H2. The number of layers of GSAM is set to L3, and the number of hidden sizes is set to H3.

[0159] During the model training process, you can also use the MSE loss function for optimization, use the Adam optimizer to train SSTNet, and set its initial learning rate, model batch size, and training termination conditions.

[0160] In addition, in order to comprehensively evaluate the prediction performance of the spatiotemporal prediction model of sea ice density, the following indicators can be used: root mean square error RMSE, mean absolute error MAE, mean absolute percentage error MAPE, determination coefficient R 2 and binaryaccuracyBACC to evaluate the model's spatiotemporal prediction performance.

[0161] The specific formula is as follows:

[0162]

[0163]

[0164] Where N is the total number of pixels in a row or column of the input image, X ij is the value of the pixel at the i-th row and j-th column of the non-land grid of the real image, To predict the pixel values ​​corresponding to the non-land grids in the image, is the average value of non-land grids in the real image.

[0165] RMSE, MAPE and MAE are common indicators for error evaluation, which can directly reflect the relative or absolute error between the predicted results and the observed results. For MAPE, if X ij If the coefficient of determination is zero, the denominator can be replaced by 0.1. 2 Measuring from the perspective of correlation can more clearly reflect the similarities and differences in evolutionary trends.

[0166] BACC is mainly used to measure the proximity of the predicted image sea ice edge to the real image. Grids with SIC ≤ 15% are defined as open waters, and grids with SIC> 15% are defined as sea ice. Sea ice extent (SIE) is defined as the total area covered by grid cells with SIC> 15%. The integrated ice edge error (IIEE) is the sum of all areas where the predicted SIE is overestimated or underestimated, and the area of ​​active grid cell region is the maximum daily SIE observed from 1988 to 2022.

[0167] Combine the following Figures 7 to 9 , describing the performance evaluation and comparative analysis of the method for predicting sea ice density of the present invention in this embodiment.

[0168] Step 1: Create a training dataset.

[0169] The sea ice density data for all years from 1978 to 2022 were obtained from NSIDC. The spatial resolution of the data was 25 km, and the Arctic stereographic projection was used. In a given 25 km grid cell, the spatial resolution of each image was 448×304×1 (H×W×C), because the data values ​​of the passive microwave instrument were less than 15% were too uncertain to be used. Therefore, values ​​less than 15% were set to 0% ice, and the remaining data ranged from 15% to 100% density. The threshold of 15% was used as the critical value for the minimum sea ice density. Since the observation interval of the data from 1978 to 1988 was not one day, the data from 1988 to 2022 were selected, with a time span of 35 years. The experimental data set contains 12772 days. Figure 7 Two SIC examples for 2022 are shown, in March and September. The average SIC in the study area has obvious seasonal fluctuations, indicating that the large-scale evolution of SIC is mainly affected by the season, with the Arctic sea ice density reaching its maximum in March and its minimum in September.

[0170] Since the sea ice density predicted for 10 days is based on the sea ice density predicted for the previous 10 days, a window covering 10 days is used to stepwise slide the sea ice density sequence with a sliding step of 1 to generate input / output sequence samples. The length of each sequence is 20, the first 10 consecutive images are used as input, and the last 10 consecutive images are used as labels for supervised training.

[0171] Step 2: Use the Pytorch environment to build the SSTNet spatiotemporal prediction network model based on the decomposition idea and sparse spatiotemporal attention mechanism.

[0172] Step 3: Training of spatiotemporal prediction model for sea ice density.

[0173] The 35-year sea ice density dataset is divided into a training set, a validation set, and a test set: the first 29 years (1988-2016) are used for training, the next year (2017) is used for validation, and the last 5 years (2018-2022) are used for testing. The spatial resolution of each image is 448×304×1 (H×W×C), and the training and validation samples are 10564 and 348, respectively. The test data is 348 samples per year (349 in leap years), and the sea ice density dataset will be constructed.

[0174] The model adopts an end-to-end training method, and no pre-training is required. In this embodiment, the number of layers of the trend network is set to 4, the number of hidden sizes is set to 128, the number of layers of the encoder encoder and decoder decoder in the season network is set to 4, and the number of hidden sizes is set to 16. The number of layers of GSAM is set to 2, the number of hidden sizes is set to 128, and SSTNet is optimized using the MSE loss function.

[0175] The Adam optimizer was used to train SSTNet, with an initial learning rate of 0.001 and adjusted to 90% of the original value after every 2561 iterations. The batch size of the model was set to 4, and the training was stopped after 51220 iterations.

[0176] The experimental environment is a system equipped with an Intel Xeon Platinum 8352V 2.10GHz twelve-core processor and an NVIDIA GeForce RTX 4090 24GB memory GPU. Python 3.10.0 and PyTorch 2.1.0 versions are used for experiments.

[0177] In order to comprehensively evaluate the prediction performance of SSTNet, the following indicators are used: root mean square error RMSE, mean absolute error MAE, mean absolute percentage error MAPE, determination coefficient R 2 and binary accuracy BACC.

[0178] Step 4: Indicator Analysis

[0179] SSTNet's MAE, MAPE, RMSE and R 2 See Table 1.

[0180]

[0181] Table 1 Test results of SSTNet

[0182] The experimental results show that the annual changes in various indicators are small. The smallest MAE is 2.02 in 2021, the smallest MAPE is 5.747 in 2022, the smallest RMSE is 5.66 in 2021, and the largest R 2 is 0.98 in 2021, with average MAE, MAPE, RMSE, and R 2 They are 2.05, 5.9, 5.78, and 0.979 respectively. The SSTNet model shows good performance with small error and high accuracy in prediction.

[0183] In order to objectively evaluate the prediction ability of 10 consecutive days of sea ice density, the sliding step size of the test set was set to 10. With this setting, two adjacent test samples do not overlap in time. 35 samples were selected from the data from 2018 to 2022, that is, 350 consecutive days of sea ice density data each year, and the prediction results were obtained through SSTNet.

[0184] Figure 8 (a)-(e) show the true value of the daily non-land grid-mean sea ice density (NSIDC), the SSTNet predicted value and the MAE from 2018 to 2022.

[0185] Analyzing the images, we can see that SSTNet's predictions of the melting and freezing trends and extreme points of sea ice are highly consistent with the actual observations. The MAE ranges from 0.52% to 7.48%, and the prediction differences over the five years are small, showing a clear 10-day periodicity. This periodic change occurs mainly due to the increase in error caused by the extension of the prediction step.

[0186] Figure 8 (f) to (j) show the spatial distribution of the mean absolute error (MAE) of the sea ice concentration forecasts for the period 2018 to 2022. These graphs calculate the annual average MAE for all forecast days at each grid point. This error is generally less than 2% over most of the Arctic. Only in some sea ice edge regions does the spatial MAE increase slightly, but it remains below 6%. Therefore, SSTNet performs well overall in predicting Arctic sea ice concentration every 10 days.

[0187] To explore why the MAE in the sea ice edge area increases, this example calculates the sea ice edge based on the predicted value of sea ice density and uses the BACC metric to evaluate its proximity. Since the Arctic sea ice reaches its minimum value in September of the year, this example selects the forecast from September 8 to September 17, 2018-2022 as the evaluation period.

[0188] The results are as follows Fig. 9As shown. The observed ice edge is the black contour line, and the ice edge predicted by SSTNet is the green contour line. The ice edge predicted by SSTNet is close to the real ice edge, with an average BACC value of 93.85%, but as the prediction step increases, the predicted edge tends to expand outward. This shows that the sea ice extent (SIE) predicted by SSTNet is too large and there is a certain lag effect. This is mainly because the high-frequency small-scale evolution during the melting period has strong nonlinear characteristics and there are interpolation errors in the data. Overall, the prediction results of SSTNet are satisfactory and reliable.

[0189] Next, further combining the attached Figure 10-Figure 15 , describing the experimental comparison results based on the scheme of the present invention and other existing methods.

[0190] The method of the present invention is compared with mainstream models, including convolutional long short-term memory network ConvLSTM, deep learning model PredRNN, deep learning model PredRNN++, video prediction model MIM, deep learning model PhyDNet combined with physical knowledge, deep learning model PredNet for video prediction, deep learning model PredRNNv2, deep learning model MAU for processing object motion in video sequences, and deep learning model SimVP for video prediction. The average performance of these methods is shown in Table 2. It can be found that compared with other existing methods, the MAE / MAPE / RMSE / R 2 All experiments were conducted on the same dataset, and the excellent performance of each indicator confirmed the prediction performance of SSTNet. The experimental results are shown in Table 2:

[0191]

[0192] Table 2 Comparison results of average performance of this application method and mainstream models

[0193] This embodiment selects a set of predicted values ​​in March when the sea ice density reaches the maximum value and September when the sea ice density reaches the minimum value in a year to analyze the performance of SSTNet, and selects March 12-March 21 and September 8-September 17 in 2022.

[0194] Figure 10-12 The detailed situation from March 12 to March 21 is shown. The MAE, MAPE and R of each model on a daily basis 2 Quantitative performance comparison of Fig.10 shown.

[0195] The prediction performance of the SSTNet proposed in the present invention is significantly better than other existing methods. In particular, when it comes to long-term prediction, the model of the present invention is always better than other methods, and the gap between the indicators becomes more obvious as the number of prediction days increases.

[0196] To qualitatively evaluate the prediction performance of the proposed SSTNet, a comparative experiment was conducted on the daily predicted SIC by comparing it with the observed images to assess how well the prediction model can predict future SIC. Fig.11 The observed images and the images predicted by each model are shown.

[0197] Fig.12 The spatial distribution of the daily residuals for each of the compared models is shown, highlighting the differences between the observed and predicted images. Fig.12 In the figure, the residual in the yellow area is positive, which means that the predicted value is smaller than the true value, and the residual in the blue area is negative, which means that the predicted value is larger than the true value. Fig.11 and Fig.12 As shown in Figure 3, the SIC predicted by the SSTNet model is basically consistent with the observed image, and the residual is small.

[0198] In general, the errors of the models in predicting the central Arctic ice sheet are small, generally within 5%, and most of them are overestimated. In the predicted SIC, most of the anomalies occur in the marginal ice areas of the North Atlantic and North Pacific due to fluctuations in the ice margin.

[0199] While the models produced quantitatively and qualitatively similar outputs, the SIC images predicted by the SSTNet model were more consistent with the observed images than those of the other compared models (i.e., the geographic locations of over- and underestimated regions were similar, but the number of SSTNet errors was smaller than that of the compared models).

[0200] The following is a brief qualitative description of each day of the forecast: The images of the first two days show that the residuals of the SSTNet forecast images are within the range of (-5%, 5%) in most areas, except for the residuals near the Barents Sea at -10%. In contrast, the residuals of the edge ice areas of the forecast images of other comparison algorithms are basically (-10%, 10%). From the 5th to the 10th day, the residuals of the edge ice areas of the SSTNet forecast images gradually increase, but the residuals of most central areas remain within the range of (-5%, 5%), and the consistency of the sea ice edge areas is always better than other comparison models.

[0201] Figure 13-15 The details from September 8th to September 17th are shown. Fig.13 The quantitative performance comparison of MAE, MAPE and R2 of each model on a daily basis is shown. As in March, the prediction performance of the SSTNet proposed in this paper is significantly better than other existing methods.

[0202] Fig.14 These are the observed images in September and the images predicted by various models. Fig.15 The spatial distribution of the daily residuals of the comparison models is shown. The sea ice range in September has shrunk a lot, but as in March, most of the anomalies in the predicted images occur in the marginal ice area. The error of the central Arctic ice sheet is still small, basically within 5%. The consistency of the SSTNet predicted image is better than that of the images of other comparison models, and the residuals in the marginal ice area are relatively small.

[0203] Evaluation of the prediction model performance shows that SSTNet produces fewer overestimated and underestimated pixels and smaller residuals than the images of the comparison model.

[0204] In Table 3, the computational efficiency of the proposed SSTNet is summarized compared with other existing models. The training time and inference time reported in Table 3 are measured on a single NVIDIA GeForce RTX 4090 GPU, providing a comprehensive analysis result for evaluating the performance of the SSTNet proposed in this invention. Compared with other comparison methods, SSTNet has smaller parameters (Params), higher fitting efficiency, and smaller floating point operations per second (FLOPs) than most comparison models. Although SSTNet is not superior in terms of model throughput (FPS) compared to some models such as PredNet and SimVP, these additional computational costs are acceptable considering the improvement in prediction performance.

[0205] Comparative experiments on computational costs with other methods. All experiments are conducted on the same dataset.

[0206]

[0207] Table 3 Comparison results of the computational costs of this application method and other methods.

[0208] In order to evaluate the contribution of decomposition and GSAM to the prediction of sea ice density, ablation experiments were conducted on the dataset without decomposition or GSAM module. To ensure the fairness of the comparison, all parameter settings and experimental configurations are the same, and the relevant results are summarized in Table 4.

[0209] All experiments are conducted on the same dataset, specifically using the first 10 days of data and inferring the next 10 days of data.

[0210]

[0211] Table 4 Ablation study Experimental results of evaluating the decomposition idea and GSAM.

[0212] From the results, the model without decomposition deteriorates significantly in various indicators, which shows that the STL decomposition idea can effectively extract the seasonal signal and trend information of the data. In addition, it can be observed that when the model does not have GSAM, the indicators decay significantly, proving that GSAM is conducive to capturing the spatiotemporal dependence of data. In summary, the ablation experiment proves that both the decomposition idea and the GSAM module are helpful to obtain satisfactory model performance.

[0213] Corresponding to the above method provided by the present invention, Fig.16 A schematic diagram of the structure of a device for spatiotemporal prediction of sea ice density provided in an embodiment of this specification is shown. The device includes:

[0214] A construction module is used to obtain the measured data of sea ice density in the target area during a preset historical period. The spatiotemporal prediction model of sea ice density is constructed based on the global sparse attention mechanism and convolutional neural network, and is trained using the historical measured data of sea ice density in the target area.

[0215] In some embodiments, it is also used to obtain measured data on sea ice density in a target area in a preset historical period and predicted data on sea ice density in a first future period.

[0216] The sea ice concentration data of the preset historical period and the sea ice concentration prediction data of the first future period are combined into sea ice concentration time data according to time;

[0217] Among them, the sea ice density prediction data for the first future period is the output result obtained by inputting the measured sea ice density data for a preset historical period into the trained sea ice density spatiotemporal prediction model.

[0218] The spatiotemporal prediction model of sea ice density is built based on the global sparse attention mechanism and convolutional neural network.

[0219] A sliding time window is used to slide on the sea ice concentration time data to obtain at least one sea ice concentration time data series.

[0220] The sea ice concentration time data can be slid based on the set sliding window length, and the step length of each sliding can be a time period.

[0221] For example, if the sea ice density for the next 10 days is to be predicted, the sea ice density data for the previous 8 days' historical period is needed. The next 10 days and the previous 8 days' historical period are taken as a time series, the length of the sliding window is set to 8, and the sliding step is 1. Then, when predicting the sea ice density for the first day, the sliding window range is the previous 8-day historical period, and the data set corresponding to the sliding window range is the sea ice density for the previous 8-day historical period. When predicting the sea ice density for the second day, the sliding window range is the previous 8-day historical period, which is the historical period from the 2nd day to the 8th day and the predicted first day. The data set corresponding to the sliding window range is the sea ice density for the previous 2nd day to the 8th day and the predicted sea ice density for the predicted first day.

[0222] The above approach meets the needs of sea ice density prediction for information at multiple historical moments and information on the current prediction time series, while limiting the length of the input time series, overcoming the time dependency problem, and avoiding the gradient vanishing or exploding phenomenon in deep learning models.

[0223] In addition, according to a specific implementation, the data set corresponding to the sliding window range can be divided into input data for prediction and data for supervised training.

[0224] The sea ice density data of a preset historical period is processed to obtain a historical sea ice density data sequence, and the historical sea ice density data sequence is input into a trained sea ice density spatiotemporal prediction model to obtain sea ice density prediction data for the first future period.

[0225] Specifically, the earliest time period of the first future time period and the latest time period of the preset historical time period are separated by one cycle.

[0226] At the same time, the historical measured data of sea ice density in the historical period and the predicted data of sea ice density in the future period are taken as a time data series, wherein the historical period can be the historical period before the earliest period in the future period.

[0227] The sliding window slides in the time data series based on the set sliding window length. The step size of each sliding can be a prediction cycle. The sliding window can include historical measured data for all or part of the historical period and all or part of the predicted data for the future period.

[0228] Each time a cycle is predicted, all data in the sea ice density time data series corresponding to each sliding window range are obtained. When predicting the next cycle, the sliding window is moved according to the step size to keep the length of the time data series input into the model each time controllable.

[0229] The processing module is used to obtain a seasonal characteristic map and a trend characteristic map of the target area based on the measured sea ice density data of the target area in a preset historical period and the sea ice density spatiotemporal prediction model. The seasonal characteristic map is used to represent the seasonal characteristics of the sea ice in the target area changing with the seasonal laws in the preset historical period, and the trend characteristic map is used to represent the spatial change trend of the sea ice in the target area in the preset historical period.

[0230] In some embodiments, the processing module is specifically used to combine the sea ice density data of the preset historical period and the sea ice density prediction data of the first future period according to time to form the sea ice density time data;

[0231] A sliding time window is used to slide on the sea ice concentration time data to obtain at least one sea ice concentration time data series.

[0232] Specifically, the sea ice density time data may be slid based on a set sliding window length, and the step length of each sliding may be a prediction cycle. Therefore, the sliding window may include all or part of the historical period and part of the first future period.

[0233] The seasonal characteristic map is used to represent the seasonal regularity of sea ice in the target area, and the trend characteristic map is used to represent the trend characteristics of sea ice in the target area.

[0234] In some embodiments, a pooling processing unit is also included for performing average pooling processing on the sea ice density data set to obtain a trend component, and removing the trend component from the sea ice density data set to obtain a seasonal component.

[0235] Specifically, for a sea ice density dataset of length L AvgPool(·) is used for moving average, that is, the corresponding elements of the two pictures are averaged to obtain X t ,

[0236] X t =AvgPool(X)

[0237] X s =XX t

[0238] X s represents the seasonal component and X t Represents the trend component.

[0239] In some embodiments, the spatiotemporal prediction model of sea ice density includes a seasonal network, and the seasonal network is used to extract a trend characteristic map of sea ice changing with seasonal laws;

[0240] The processing module is specifically used to obtain the seasonal characteristic map of the target area based on the measured sea ice density data of the target area in a preset historical period and the sea ice density spatiotemporal prediction model, which specifically includes:

[0241] Extracting seasonal components from measured sea ice density data, where the seasonal components are used to represent the seasonal regularity of the measured sea ice density data;

[0242] The seasonal component is input into the seasonal network to obtain the seasonal feature map.

[0243] In some other embodiments, the processing module is specifically used to extract a seasonal component from the first sea ice concentration time data sequence, and the seasonal component is used to represent a seasonal regular part in the sea ice concentration time data sequence;

[0244] The seasonal component is input into the seasonal network to obtain the seasonal feature map.

[0245] In some embodiments, the sea ice density spatiotemporal prediction model includes a trend network, and the trend network is used to obtain a characteristic map of the spatial variation trend of sea ice;

[0246] The processing module is specifically used to obtain a trend characteristic map of the target area based on the measured sea ice density data of the target area in a preset historical period and the sea ice density spatiotemporal prediction model, which specifically includes:

[0247] Extracting a trend component from the measured sea ice density data, the trend component is used to represent the trend change part in the measured sea ice density data;

[0248] The seasonal component is input into the seasonal network to obtain the seasonal feature map.

[0249] In some other embodiments, the processing module is specifically used to extract a trend component from the first sea ice concentration time data series, the trend component being used to represent a trend change portion in the sea ice concentration time data series;

[0250] The seasonal component is input into the seasonal network to obtain the seasonal feature map.

[0251] In some embodiments, the season network includes an encoder, a global sparse attention mechanism GSAM, and a decoder.

[0252] The encoder includes at least one convolutional layer for encoding a high-dimensional frame of the sea ice density dataset into a low-dimensional latent space, and the GSAM includes at least one spatiotemporal attention mechanism for learning latent spatiotemporal information in the sea ice density dataset from the low-dimensional latent space;

[0253] The decoder comprises at least one transposed convolutional layer for decoding the potential spatiotemporal information in the seasonal component to obtain a seasonal feature map;

[0254] Accordingly, the implementation of this module may also include the following modules:

[0255] A first sub-processing module is used to input the seasonal component into an encoder for encoding to obtain a first feature map of a low-dimensional latent space;

[0256] A second sub-processing module, configured to input the first feature map into the GSAM to obtain a second feature map containing potential spatiotemporal information;

[0257] Further, in some more specific embodiments, GSAM includes a channel attention mechanism and a spatial attention mechanism. The channel attention mechanism includes a sparse multilayer perceptron sMLP for compressing the spatial dimension of the first feature map. The spatial attention mechanism includes at least one convolutional layer for performing spatial information fusion on the feature map output by the channel attention mechanism.

[0258] Correspondingly, the second sub-processing module is specifically used to input the first feature map into the channel attention mechanism for spatial dimension compression to obtain the first convolution feature map output by the channel attention mechanism;

[0259] Multiplying the first feature map and the first convolution feature map to obtain a first input feature map;

[0260] Input the first input feature map into the spatial attention mechanism for spatial information fusion to obtain the second convolution feature map;

[0261] Multiply the second convolution feature map and the first input feature map to obtain a second feature map.

[0262] The third sub-processing module is used to input the second feature map into the decoder for decoding to obtain a seasonal feature map.

[0263] In other embodiments, the trend network includes at least one convolutional layer for extracting trend features in the trend component.

[0264] Accordingly, the implementation of this module may also include the following module fourth sub-processing module, which is used to utilize Get trend characteristics Figure X TN ;

[0265] in, is the trend component X t , is the trend component obtained by the i-th round of convolution, σ is the nonlinear activation function, GroupNorm is the normalization layer, and N T is the number of convolutional layers in the trend network, and Conv2d is the convolutional layer.

[0266] The fusion module is used to obtain the sea ice density prediction data of the target area in the future period based on the seasonal characteristic map and the trend characteristic map.

[0267] In other embodiments, the fusion module is used to predict the sea ice density prediction data of the target area in the second future period based on the seasonal characteristic map and the trend characteristic map.

[0268] In some embodiments, the implementation of this step may include:

[0269] The seasonal characteristic map and the trend characteristic map are added together to obtain the predicted data of sea ice density in the target area in the future period.

[0270] In some embodiments, based on the seasonal characteristic map and the trend characteristic map, sea ice density prediction data for the target area in a second future period is predicted.

[0271] The end time of the first future time period is earlier than the start time of the second future time period.

[0272] According to another embodiment, there is also provided a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute a combination of Figure 2 The method described.

[0273] According to another embodiment of the present invention, a computing device is provided, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the Figure 2 The method described.

[0274] Those skilled in the art should be aware that in one or more of the above examples, the functions described in the present invention can be implemented by hardware, software, firmware or any combination thereof. When implemented by software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium.

[0275] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made on the basis of the technical solution of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for spatiotemporal prediction of sea ice density, characterized in that: The method comprises: Obtain the measured data of sea ice density in the target area during a preset historical period. The spatiotemporal prediction model of sea ice density is constructed based on the global sparse attention mechanism and convolutional neural network, and is trained using the historical measured data of sea ice density in the target area. According to the measured data of sea ice density of the target area in a preset historical period and the spatiotemporal prediction model of sea ice density, a seasonal characteristic map of the target area and a trend characteristic map of the target area are obtained, wherein the seasonal characteristic map is used to represent the seasonal characteristics of the sea ice in the target area changing with the seasonal law in the preset historical period, and the trend characteristic map is used to represent the spatial variation trend of the sea ice in the target area in the preset historical period; Based on the seasonal characteristic map and the trend characteristic map, obtaining sea ice density prediction data for the target area in a future period; The sea ice density spatiotemporal prediction model includes a seasonal network, and the seasonal network is used to extract a trend characteristic diagram of sea ice changing with seasonal laws; According to the measured sea ice density data of the target area in a preset historical period and the spatiotemporal prediction model of sea ice density, a seasonal characteristic map of the target area is obtained, which specifically includes: Extracting a seasonal component from the measured sea ice density data, wherein the seasonal component is used to represent a seasonal regular part of the measured sea ice density data; Inputting the seasonal component into the seasonal network to obtain the seasonal characteristic graph; The sea ice density spatiotemporal prediction model includes a trend network, and the trend network is used to obtain a characteristic map of the spatial variation trend of sea ice; According to the measured sea ice density data of the target area in a preset historical period and the spatiotemporal prediction model of sea ice density, a trend characteristic map of the target area is obtained, which specifically includes: Extracting a trend component from the measured sea ice density data, wherein the trend component is used to represent a trend change portion in the measured sea ice density data; Inputting the trend component into the trend network to obtain the trend characteristic graph; The seasonal network includes an encoder, a global sparse attention mechanism GSAM and a decoder; The trend network includes at least one convolutional layer for extracting trend features in the trend component.

2. The method according to claim 1, characterized in that The encoder comprises at least one convolutional layer for encoding high-dimensional frames in the seasonal component into a low-dimensional latent space; The GSAM comprises at least one spatiotemporal attention mechanism for learning latent spatiotemporal information in the seasonal component from a low-dimensional latent space; The decoder comprises at least one transposed convolution layer for decoding the potential spatiotemporal information in the seasonal component to obtain the seasonal feature map; The step of inputting the seasonal component into the seasonal network to obtain the seasonal characteristic graph specifically includes: Inputting the seasonal component into the encoder for encoding to obtain a first feature map of a low-dimensional latent space; Inputting the first feature map into the GSAM to obtain a second feature map containing potential spatiotemporal information; The second characteristic map is input into the decoder for decoding to obtain the seasonal characteristic map.

3. The method according to claim 2, characterized in that The GSAM includes a channel attention mechanism and a spatial attention mechanism, wherein the channel attention mechanism includes a sparse multilayer perceptron sMLP for compressing the spatial dimension of the first feature map, and the spatial attention mechanism includes at least one convolutional layer for fusing spatial information of the feature map output by the channel attention mechanism; The step of inputting the first characteristic graph into the GSAM to obtain a second characteristic graph output by the GSAM specifically includes: Inputting the first feature map into the channel attention mechanism for spatial dimension compression to obtain a first convolution feature map output by the channel attention mechanism; Multiplying the first feature map and the first convolutional feature map to obtain a first input feature map; Inputting the first input feature map into the spatial attention mechanism to perform spatial information fusion to obtain a second convolution feature map; Multiplying the second convolutional feature map and the first input feature map to obtain the second feature map.

4. The method according to claim 1, characterized in that: The step of inputting the trend component into the trend network to obtain the trend characteristic graph specifically includes: Using the trend component , get the trend characteristic diagram ; in, is the trend component , is a nonlinear activation function, is the normalization layer, is the number of convolutional layers in the trend network, It is a convolutional layer.

5. The method according to claim 1, characterized in that The obtaining, based on the seasonal characteristic map and the trend characteristic map, predicted data of sea ice density in the target area in the future period specifically includes: The seasonal characteristic map and the trend characteristic map are added together to obtain the sea ice density prediction data of the target area in the future period.

6. A device for spatiotemporal prediction of sea ice density, characterized in that: The device comprises: A construction module is used to obtain the measured data of sea ice density in the target area in a preset historical period, wherein the spatiotemporal prediction model of sea ice density is constructed based on the global sparse attention mechanism and convolutional neural network, and is trained using the historical measured data of sea ice density in the target area; A processing module, for obtaining a seasonal characteristic map and a trend characteristic map of the target area according to the measured sea ice density data of the target area in a preset historical period and the spatiotemporal prediction model of the sea ice density, wherein the seasonal characteristic map is used to represent the seasonal characteristics of the sea ice in the target area changing with the seasonal law in the preset historical period, and the trend characteristic map is used to represent the spatial variation trend of the sea ice in the target area in the preset historical period; A fusion module, used for obtaining the sea ice density prediction data of the target area in the future period based on the seasonal characteristic map and the trend characteristic map; The sea ice density spatiotemporal prediction model includes a seasonal network, and the seasonal network is used to extract a trend characteristic diagram of sea ice changing with seasonal laws; According to the measured sea ice density data of the target area in a preset historical period and the spatiotemporal prediction model of sea ice density, a seasonal characteristic map of the target area is obtained, which specifically includes: Extracting a seasonal component from the measured sea ice density data, wherein the seasonal component is used to represent a seasonal regular part of the measured sea ice density data; Inputting the seasonal component into the seasonal network to obtain the seasonal characteristic graph; The sea ice density spatiotemporal prediction model includes a trend network, and the trend network is used to obtain a characteristic map of the spatial variation trend of sea ice; According to the measured sea ice density data of the target area in a preset historical period and the spatiotemporal prediction model of sea ice density, a trend characteristic map of the target area is obtained, which specifically includes: Extracting a trend component from the measured sea ice density data, wherein the trend component is used to represent a trend change portion in the measured sea ice density data; Inputting the trend component into the trend network to obtain the trend characteristic graph; The seasonal network includes an encoder, a global sparse attention mechanism GSAM and a decoder; The trend network includes at least one convolutional layer for extracting trend features in the trend component.

7. A computer storage medium, characterized in that: The computer storage medium stores a computer program, and when the computer program is executed by one or more processors, the method for spatiotemporal prediction of sea ice density according to any one of claims 1 to 5 is implemented.

8. An electronic device, characterized in that: The method comprises a memory and one or more processors, wherein a computer program is stored in the memory, and when the computer program is executed by the one or more processors, the method for spatiotemporal prediction of sea ice density as claimed in any one of claims 1 to 5 is implemented.

Images